Mental model
Approval is a gate over authority, not a generic confirmation dialog. Its strength should rise with consequence, irreversibility, uncertainty, affected scope, and the difficulty of detecting harm after execution.
Theory
Classify actions before runtime: read-only, reversible write, consequential write, privileged administration, or prohibited. The application, not the model, maps that classification to an approval policy. Place the gate after arguments and affected resources are known but before side effects occur.
An approval request should identify the goal, exact operation, target, important parameters, expected effect, uncertainty, alternatives, and rollback path. The resulting authorization token must be short-lived and bound to the user, operation, arguments, target, and current state. Changing any material field invalidates approval.
Alternatives and trade-offs
Automatic execution is appropriate for low-consequence, reversible actions under strict limits. Batch approval reduces interruption but increases scope. Policy-based preauthorization works for repetitive operations with narrow parameters. A human checkpoint improves control only when the reviewer has time, competence, and real ability to reject.
Failure modes and misconceptions
Approval fatigue turns prompts into decoration. Asking before parameters are known is uninformed consent. Reusing a broad approval token enables confused-deputy behavior. Hiding model uncertainty or tool provenance prevents meaningful review. Human involvement cannot compensate for excessive underlying permissions.
Knowledge check
Which fields must an approval token bind so the executed operation cannot drift from what the reviewer saw?
Decision scenario
A database agent may generate a migration and run a read-only plan automatically. Applying the migration requires a reviewer to see the target database, statements, lock estimate, affected tables, backup state, and rollback command. Approval expires if the plan changes.
Learning outcomes
- Explain Human Approval Boundaries as a system mechanism rather than a slogan.
- Compare its alternatives, trade-offs, and production failure modes.
- Apply the concept to a decision and identify evidence that would validate it.
Trade-offs
Using Human Approval Boundaries can improve capability or control, but it also introduces cost, latency, complexity, and failure modes that must be measured against an explicit objective.